custom_features

Back to custom extensions

Use custom feature functions when a transformation belongs after preprocessing and before model fitting. The output must be a numeric feature matrix aligned to the input date index.

Direct Function

mf.feature_engineering.custom_features(
    data,
    func,
    *,
    columns=None,
    name=None,
    prefix=None,
    **params,
) -> pandas.DataFrame

Callable Signature

func(source: pandas.DataFrame, *, metadata: dict, **params)

Accepted Return Types

Return type	Requirement
DataFrame	Same date index or same row count as source.
Series	Same date index or same row count as source.
1-D array-like	Length equals len(source).
2-D array-like	Row count equals len(source).

The returned feature table receives attrs["macroforecast_metadata_schema"] and attrs["macroforecast_metadata"].

Runner-Safe Step

mf.feature_engineering.custom_step(
    name,
    func=None,
    *,
    fit_func=None,
    transform_func=None,
    columns=None,
    requires_target=False,
    prefix=None,
    **params,
) -> dict

Stateless Step

def square_feature(source, *, metadata=None, suffix="sq"):
    column = source.columns[0]
    return pandas.DataFrame(
        {f"{column}_{suffix}": source[column] ** 2},
        index=source.index,
    )

features = mf.feature_engineering.feature_spec(
    target="target",
    horizon=1,
    predictors=["x", "z"],
    steps=[
        mf.feature_engineering.custom_step(
            "x_square",
            square_feature,
            columns=["x"],
        ),
    ],
)

Fitted Step

features = mf.feature_engineering.feature_spec(
    target="target",
    horizon=1,
    predictors="all",
    steps=[
        mf.feature_engineering.custom_step(
            "my_factor",
            fit_func=my_factor_fit,
            transform_func=my_factor_transform,
            columns=["PAYEMS", "UNRATE", "HOUST"],
            requires_target=True,
            prefix="myf",
            n_components=2,
        ),
    ],
)

Callable	Contract
fit_func	fit_func(source, target=None, metadata=None, **params) -> state
transform_func	transform_func(source, state=state, metadata=None, **params) -> feature output
fitted state object	state.transform(source) -> feature output
state-aware func	func(source, state=state, metadata=None, **params) -> feature output

Set requires_target=True only when the feature fit step needs the resolved target. Transform-time code should not use future target values.

Flow

feature_set = features.fit_transform(processed.panel)
fit = mf.models.ridge(feature_set.X.dropna(), feature_set.y.iloc[:, 0].dropna())